Behavioural and neuroanatomical correlates of auditory speech analysis in primary progressive aphasias

Ten patients with nfvPPA (five females; mean age 71.2 ± 8.9 (SD) years) and nine patients with svPPA (three females; mean age 63.8 ± 4.6 years) were recruited consecutively via a specialist cognitive clinic. All patients fulfilled current consensus criteria for a probable or definite diagnosis of the relevant PPA syndrome [1] and this was corroborated by general neuropsychological assessment and brain MRI findings. No patient had radiological evidence of significant co-morbid cerebrovascular disease. Nineteen healthy older individuals (10 females; mean age 69.4 ± 4.5 years) with no history of neurological or psychiatric illness also participated. No participant had a history of clinically significant hearing loss; peripheral hearing function was assessed in all participants using pure tone audiometry (details shown in Additional file 1). Demographic, clinical and neuropsychological data for all participants are summarised in Table 1.

All participants gave informed consent. Ethical approval for the study was granted by the National Hospital for Neurology and Neurosurgery and the University College London Research Ethics Committees, in accordance with the Declaration of Helsinki.

For the experimental stimuli, we created sequences of spoken syllables consisting of consonant–vowel or vowel–consonant phoneme combinations. We chose the syllables ‘af’, ‘ba’, ‘da’, ‘mo’, ‘om’, ‘or’, ‘po’ and ‘ro’ for their high intelligibility and identifiability, based on pilot work in five young adult listeners (details shown in Additional file 1). Syllables were recorded in a standard southern English accent by a young adult male speaker. Using MATLAB R2012a (https://​uk.​mathworks.​com/​), syllables were concatenated with random ordering to form sequences each comprising 20 syllables of duration 240 ms and fundamental frequency 100 Hz, with intervening silent intervals. The overall sequence duration (7.65 seconds) and root mean square intensity were fixed across sequences. Different conditions were created by independently varying three sequence parameters: temporal regularity, phonemic structure and entropy (fundamental information content, a measure of signal unpredictability).

Temporal regularity was varied by altering the inter-syllabic interval such that this was either kept constant at 150 ms (isochronous condition) or randomly allocated in the range 50–250 ms around a mean of 150 ms (anisochronous condition) while maintaining the same overall sequence tempo. Phonemic structure was varied using a previously described procedure of spectral rotation [29]: this manipulation preserves overall acoustic spectro-temporal complexity and bandwidth but profoundly affects spectral detail, by inverting the acoustic frequency spectrum and thereby rendering the rotated signal unintelligible as human speech (listeners generally describe it as ‘alien’ or ‘computer speech’). We synthesised stimulus conditions in which the constituent syllables comprising each sequence were either all unrotated (natural) or all spectrally rotated (unintelligible). Speech signal predictability was varied as an index of fundamental information content or entropy of the syllable sequences: in classical information theory, signals with high fundamental information content (or entropy) have low predictability. We adapted a previously described procedure [21] to manipulate the overall predictability of the pitch contour of each syllable sequence. This procedure varied the fundamental frequency (pitch) of constituent syllables over a half-octave range, using a 20-note octave division that did not conform to Western musical intervals; pitch sequences (prosodic contours) were based on inverse Fourier transforms of f ⁿ power spectra with values n = 0 (no correlation between consecutive syllable pitch values; the low signal predictability–high entropy condition) and n = 4 (high correlation between consecutive syllable pitch values, approaching a sine-wave contour; the high signal predictability–low entropy condition). It is important to note that this prosodic manipulation does not correspond to any single feature of natural prosody: rather, it taps into a generic statistical property of prosodic contours (the correlation structure of the syllabic pitch sequence) that is potentially relevant to many kinds of patterns in speech signals.

The stimuli are schematised in Fig. 1; examples are provided in Additional files 2, 3, 4, 5, 6 and 7.

These experimental stimuli formed the basis for three two-alternative, forced-choice psychoacoustic tasks, each probing a particular dimension of auditory processing. Separate tests were administered to assess pitch pattern analysis (predictable vs unpredictable sequences), temporal processing (regular vs irregular sequences) and phoneme detection (natural vs artificial (spectrally rotated, unintelligible) phonemes). Tests were administered in the same order to all participants: first, pitch pattern analysis; second, temporal processing; and third, phoneme detection. For each test, 20 stimuli (10 representing each of the two conditions of interest) were presented. For the test assessing processing of prosodic predictability, participants were asked to decide whether the sounds were arranged randomly or following a pattern; for the test assessing temporal processing, participants were asked on each trial to decide whether the sounds they heard came regularly or irregularly; and for the test assessing processing of phonemic structure, participants were asked to decide whether the sounds were made by a human or by a computer. Pictorial cue cards (see Additional file 8) were used as tools to ensure understanding of the task instructions in practice trials, before commencement of the test proper. On each trial, participants could respond verbally or by pointing to the relevant card.

Stimuli were presented in randomised order via a notebook computer running the Cogent v1.32 extension of MATLAB (www.​vislab.​ucl.​ac.​uk/​cogent_​2000.​php). No feedback about task performance was given and no time limits were imposed. Participant responses were recorded for offline analysis.

Clinical and behavioural data were analysed using Stata® v14.1. Participant groups were compared on demographic and other clinical variables using two-tailed, two-sample t tests for continuous variables and chi-square tests for categorical variables. Non-parametric Mann–Whitney U tests were used to compare groups on neuropsychological parameters where residuals were non-normally distributed.

In order to compare groups for peripheral hearing function, we first created a composite pure tone average score based on the average volume (dB) required for tone detection at 500, 1000 and 2000 Hz, for each ear separately. Using data from the best ear for each participant, scores within the range of 0–25 dB were categorised as ‘normal’, scores of 26–40 dB were classified as ‘mild hearing loss’ and scores of 41–55 dB classified as ‘moderate hearing loss’. Based on these classifications, each participant’s hearing function was treated as a categorical variable and Fisher’s exact test was used to compare groups.

In separate regression (Spearman’s rank correlation) analyses over the participant cohort, we assessed experimental psychoacoustic task performance against background executive function (WASI Matrices score; a proxy for disease severity) and a standard measure of phoneme discrimination (PALPA-3 score).

Volumetric brain MR images were acquired for all patients in a 3 Tesla Siemens Tim Trio MRI scanner, using a 32-channel receiver array head coil and a T1-weighted sagittal 3D magnetisation prepared rapid gradient echo (MPRAGE) sequence (TE = 2.9 ms, TI = 900 ms, TR = 2200 ms), with dimensions 256 mm × 256 mm × 208 mm and voxel size 1.1 mm × 1.1 mm × 1.1 mm.

For the VBM analysis, patients’ brain images were first pre-processed and normalised to MNI space using SPM12 software (http://​www.​fil.​ion.​ucl.​ac.​uk/​spm/​software/​spm12/​) and the DARTEL toolbox with default parameters running under MATLAB R2012a. Images were smoothed using a 6-mm full-width at half-maximum Gaussian (FWHM) kernel. To control for individual differences in total (pre-morbid) brain size, total intracranial volume was calculated for each participant by summing white matter, grey matter and cerebrospinal fluid volumes post segmentation. A study-specific average brain upon which to overlay statistical parametric maps was created by warping all patients’ native-space whole-brain images to the final DARTEL template and using the ImCalc function to generate an average of these images.

We assessed neuroanatomical correlates of experimental behavioural task performance in a separate analysis. Voxel intensity was modelled for the combined patient cohort as a function of performance on each of the experimental psychoacoustic tasks in a multiple regression design incorporating age, total intracranial volume, disease duration and group membership as nuisance covariates. An explicit brain mask was created using an automatic mask-creation strategy described previously [30]. Statistical parametric maps were thresholded at a peak-level p < 0.05 after family-wise error (FWE) correction for multiple voxel-wise comparisons within a pre-defined region of interest, based on neuroanatomical predictions from previous studies. Correlates of behavioural performance on the temporal regularity and prosodic predictability tests were assessed within a region comprising the bilateral posterior superior temporal gyrus, planum temporale, supramarginal gyrus, supplementary motor area, anterior cingulate and striatum [20‐23]. Grey matter correlates of performance on the phoneme detection test were assessed with a more restricted sub-region comprising the left posterior superior temporal gyrus, planum temporale and supramarginal gyrus [25‐28]. Anatomical regions were derived from Oxford–Harvard cortical maps [31] and are depicted in Additional file 9.